Search for: All records

Theory of mind, the ability to model others’ thoughts and desires, is a cornerstone of human social intelligence. This makes it an important challenge for the machine learning community, but previous works mainly attempt to design agents that model the "mental state" of others as passive observers or in specific predefined roles, such as in speaker-listener scenarios. In contrast, we propose to model machine theory of mind in a more general symmetric scenario. We introduce a multi-agent environment SymmToM where, like in real life, all agents can speak, listen, see other agents, and move freely through the world. Effective strategies to maximize an agent’s reward require it to develop a theory of mind. We show that reinforcement learning agents that model the mental states of others achieve significant performance improvements over agents with no such theory of mind model. Importantly, our best agents still fail to achieve performance comparable to agents with access to the gold-standard mental state of other agents, demonstrating that the modeling of theory of mind in multi-agent scenarios is very much an open challenge. 

Children do not learn language from passively analyzing correlations between language and observations, but from interaction with caregivers or peers. The non-nativist approach claims that the main driver of language learning should be to achieve communicative goals. Imitation, on the other hand, is another natural desire that many argue influences language learning. However, there are still gaps in the research on what roles communicative goals and imitating linguistic input play in language acquisition, due to the difficulty of performing comprehensive experiments with human learners. In this paper, we propose a computational framework using simulated experiments that allows us to compare the roles of the two drivers. Specifically, we simulate a two-way communication game between a speaker, corresponding to a language learner, and a listener, corresponding to a caregiver or teacher. The speaker's communicative goals are modeled as rewards for successful completion of a referential game, and imitation is performed by mimicking feedback from the listener. The listener adaptively chooses to give feedback and makes choices based on the speaker's utterances. With empirical results on naturalistic visual and language data, we find that communicative goals play an important role in driving language learning, whereas imitation accelerates the learning process. We also find that (1) models trained with communicative goals tend to use minimal vocabulary and utterances and overextend them to concepts outside the original word meanings; (2) the strategy with which the listener provides feedback also influences the learning results and speed. Code and data for replicating the experiments are available (https://bit.ly/interactgym) to spur future research on models for computational studies of language learning. 

Recent progress in natural language generation has raised dual-use concerns. While applications like summarization and translation are positive, the underlying technology also might enable adversaries to generate neural fake news: targeted propaganda that closely mimics the style of real news. Modern computer security relies on careful threat modeling: identifying potential threats and vulnerabilities from an adversary's point of view, and exploring potential mitigations to these threats. Likewise, developing robust defenses against neural fake news requires us first to carefully investigate and characterize the risks of these models. We thus present a model for controllable text generation called Grover. Given a headline like `Link Found Between Vaccines and Autism,' Grover can generate the rest of the article; humans find these generations to be more trustworthy than human-written disinformation. Developing robust verification techniques against generators like Grover is critical. We find that best current discriminators can classify neural fake news from real, human-written, news with 73% accuracy, assuming access to a moderate level of training data. Counterintuitively, the best defense against Grover turns out to be Grover itself, with 92% accuracy, demonstrating the importance of public release of strong generators. We investigate these results further, showing that exposure bias -- and sampling strategies that alleviate its effects -- both leave artifacts that similar discriminators can pick up on. We conclude by discussing ethical issues regarding the technology, and plan to release Grover publicly, helping pave the way for better detection of neural fake news.